Draft version November 4, 2018

Preprint typeset using LATEX style emulateapj v. 04/03/99

BEST UNBIASED ESTIMATORS FOR THE THREE-POINT CORRELATORS OF THE COSMIC

MICROWAVE BACKGROUND RADIATION
Alejandro Gangui1,2,3 and Jérôme Martin3
1 Instituto de Astronomı́a y Fı́sica del Espacio, Ciudad Universitaria, 1428 Buenos Aires, Argentina
2 Dept. de Fı́sica, Universidad de Buenos Aires, Ciudad Universitaria – Pab. 1, 1428 Buenos Aires, Argentina
3 DARC, Observatoire de Paris–Meudon, UMR 8629 CNRS, 92195 Meudon Cedex, France.
arXiv:astro-ph/0001361v2 22 Feb 2001

E-mail: gangui@irupé.obspm.fr, martin@edelweiss.obspm.fr

Draft version November 4, 2018

ABSTRACT
Measuring the three-point correlators of the Cosmic Microwave Background (CMB) anisotropies could
help to get a handle on the level of non-Gaussianity present in the observational datasets and therefore
would strongly constrain models of the early Universe. However, typically, the expected non-Gaussian
signal is very small. Therefore, one has to face the problem of extracting it from the noise, in particular
from the ‘cosmic variance’ noise. For this purpose, one has to construct the best unbiased estimators
for the three-point correlators that are needed for concrete detections of non-Gaussian features. In this
article, we study this problem for both the CMB third moment and the CMB angular bispectrum. We
emphasize that the knowledge of the best estimator for the former does not permit one to infer the best
estimator for the latter and vice versa. We present the corresponding best unbiased estimators in both
cases and compute their corresponding cosmic variances.
Subject headings: cosmic microwave background — methods: analytical — cosmology: theory — large
scale structure of universe — early Universe
1. INTRODUCTION

1
The Cosmic Microwave Background (CMB) has been recognized as one of the best tools for studying the early Universe
(e.g. Scott 1999). In particular, the statistical properties of the CMB anisotropies are a powerful means to discriminate
amongst the possible scenarios. This is because, in general, different models predict different statistical properties. For
example, the simplest models of inflation predict that the temperature anisotropies should obey a Gaussian statistics and
therefore any non-vanishing measurement of a three-point correlator (in a sense to be precised below) would automatically
ruled out such models, a very interesting result indeed.
From a practical point of view, measuring any non-Gaussianity in the data is a very difficult task since the signal is
typically very small. Of course, this signal should be compared to the noise and what really matters is the signal to noise
ratio. The noise can have many different origins including instrumental errors, foregrounds contamination or incomplete
sky coverage. Another source of error is the so-called ‘cosmic variance’. Roughly speaking, it comes from the fact that we
only have access to one realization of the temperature anisotropies whereas theoretical predictions are expressed through
ensemble averages. In a Gaussian model, for example, the mean value of any three-point correlator has to vanish but this
does not guarantee that a concrete detection of a non-zero signal on the sky would be in contradiction with the model
(Scaramella & Vittorio 1991, Srednicki 1993). The important point is that the cosmic variance can dominate the other
sources of error, as this is in fact the case for the two-point correlators on large angular scales. Therefore, if one wants to
1 A.G. would like to dedicate this article to the memory of his Ph.D. supervisor, Dennis William Sciama.
1
2

unveil non-Gaussianity, it is necessary to address the cosmic variance problem for the three-point correlators. The usual
way to deal with this problem is to construct estimators by performing spatial averages on the celestial sphere and to
find the one which has the smallest possible variance. The aim of this paper is then to find the best unbiased estimators
both for the third moment am

1 m2 m3 
ℓ 1 aℓ 2 aℓ 3 and for the angular bispectrum Cℓ1 ℓ2 ℓ3 and to display the corresponding cosmic
variances.
Recently, there has been a lot of activity in the subject triggered by the finding that non-Gaussianities are present in
the 4-yr COBE-DMR data (Ferreira et al. 1998, Pando et al. 1998). Further analyses have confirmed this result (e.g.
Bromley & Tegmark 1999). However, soon after, it was demonstrated by Banday et al. (1999) that the non-Gaussian
signal is driven by the 53 GHz data. This systematic artifact in the CMB maps rejects a possible cosmological origin.
More generally, it is clear that the presence of foregrounds (Bouchet & Gispert 1999, Tegmark et al. 1999) renders difficult
the detection of a genuine non-Gaussian signal. Nevertheless, one should expect non-Gaussian features to be present in
the CMB anisotropy datasets. These could be produced in the early Universe during inflation either because the initial
conditions are non-Gaussian themselves [i.e. the quantum initial state is not the vacuum (Martin et al. 1999, Contaldi et
al. 1999)] or owing to the existence of couplings between different perturbation modes at the non-linear level (Gangui et
al. 1994, Gangui 1994, Linde & Mukhanov 1997). In the context of slow-roll inflation, the CMB bispectrum has recently
been studied in (Gangui & Martin 2000, Wang & Kamionkowski 2000). Even if non-Gaussianities are not primordial
in origin, they will nevertheless arise during later stages of evolution. In this context, the Rees-Sciama effect will build
up a small but non-vanishing signal (Luo & Schramm 1993, Mollerach et al. 1995, Munshi et al. 1995). Also, cosmic
topological defects of the vacuum, like strings and textures, are amongst the best motivated sources for non-Gaussian
features (Bouchet et al. 1988, Ferreira & Magueijo 1997, Avelino et al. 1998, Gangui & Perivolaropoulos 1995, Gangui &
Mollerach 1996). Regarding secondary sources, Goldberg & Spergel (1999) and Spergel & Goldberg (1999) have recently
calculated the angular bispectrum due to second order gravitational effects like the correlation of lensing of CMB photons
and secondary anisotropies coming from the Integrated Sachs Wolfe effect and thermal Sunyaev-Zel’dovich effect. In
the same line, Cooray and Hu (1999) have taken into account further additional contributions to the bispectrum in the
presence of reionization. Other approaches to the study of non-Gaussian features include preferred-direction statistics
for sky maps (Bunn & Scott 1999), the three-point correlation function (Falk et al. 1993, Hinshaw et al. 1994, Gangui
et al. 1994, Hinshaw et al. 1995), lensing statistics (Bernardeau 1997, Winitzki 1998, Zaldarriaga 1999), the genus and
Euler-Poincaré statistics (Coles 1989, Gott et al. 1990, Smoot 1999), peak statistics (Bond & Efstathiou 1987, Kogut et
al. 1995, Kogut et al. 1996), correlation function of peaks (Heavens & Sheth 1999), Minkowski functionals (Winitzki &
Kosowsky 1998) and wavelet analyses (Popa 1998, Hobson et al. 1998).
This article is organized as follows. In the next section, the general strategy for finding best estimators is exposed. As
a warm up, in the third section, we implement this strategy for the two-point correlators. The fourth section is the core
of the article. There we explicitly derive, for the first time, the best unbiased estimator for the angular bispectrum and
show its corresponding variance. Except for an overall normalization factor, this estimator turns out to be the one already
employed by Ferreira et al. (1998) and other authors recently. Our result places their choice on a firm basis. Next, we
find the expression for the best unbiased estimator for the third moment. An earlier study was performed in (Heavens
1998); however, our findings go beyond the results obtained in that article and, moreover, are explicit. Moreover, we
present its corresponding variance. In addition, we also emphasize that the knowledge of the best estimator for the third
moment does not allow one to infer the best estimator for the angular bispectrum and vice versa. In the last section, we
briefly present our main conclusions. We finish up with a short Appendix which includes formulae related to the inverse
two-point correlation function.
 3

2. GENERAL STRATEGY FOR FINDING THE BEST ESTIMATOR

In this section, we expose the cosmic variance problem from the viewpoint of the theory of cosmological perturbations of
quantum-mechanical origin and describe the method of the best unbiased estimators. This theory rests on the principles
of general relativity and quantum field theory. At the beginning of the inflationary phase (Guth 1981, Linde 1983a,
Albrecht & Steinhardt 1982, Linde 1983b) the Friedmann-Lemaı̂tre-Robertson-Walker background spacetime already
behaves classically whereas the excitations of the metric around this background are still quantum mechanical in nature.
Technically, this means that the perturbed metric must be considered as a quantum operator. This operator either
represents density perturbations or gravitational waves. In each case, the quantization can be carried out in a consistent
way (Mukhanov & Chibisov 1981, Hawking 1982, Starobinsky 1982, Bardeen et al. 1983, Mukhanov et al. 1992, Grishchuk
1993, Martin & Schwarz 1998, 1999). Then, the (zero-point) quantum fluctuations, which are the seeds of the cosmological
perturbations, are amplified during inflation owing to the particle-creation phenomenon or squeezing effect (Grishchuk
& Sidorov 1990). Next, these primordial fluctuations give rise to the large scale structures and to the CMB anisotropies
observed today in our Universe.
One should also discuss the choice of the quantum state in which the metric operator is placed. Obviously, it is not
possible to prepare the initial state of the Universe and therefore the choice of the quantum state of the perturbations is
a priori free unless some theory of the initial conditions is provided [for example, quantum cosmology (Halliwell 1989)].
Usually, it is assumed that the initial state is the vacuum although different hypothesis are possible (Brandenberger &
Hill 1986, Martin et al. 1999, Contaldi et al. 1999). If the initial state is the vacuum, then the corresponding statistical
properties are Gaussian. This is because the ground-state wave function of an harmonic oscillator is a Gaussian. It is
possible to avoid this general conclusion either by considering non-linear cosmological perturbations or by assuming that
the initial state is a non-vacuum state. We have recently investigated the first possibility in (Gangui & Martin 2000). The
second possibility has been studied by Martin et al. (1999). In the latter case, non-Gaussianity is likely to be significant
only for relatively small angular scales.
It should be emphasized that the mechanism described previously is deeply rooted in the quantum-mechanical nature
of the gravitational field. The observable quantities calculated in this framework are always proportional to the Planck
length. In other words, if observations confirm the full set of inflationary predictions then the fact that δT /T 6= 0 would
be a direct observational consequence of quantum gravity.
The quantum-mechanical origin of the anisotropies in the framework of inflation raises also profound problems of
interpretation. One should not think that these problems are purely theoretical. On the contrary, they have consequences
with regards to the experimental strategy that one should follow in order to extract as much informations as possible
from the data. The fluctuations in the CMB effective temperature are linked to the perturbed metric as shown for the
first time by Sachs and Wolfe. Therefore, the fact that the perturbed metric is an operator implies that the primordial
fluctuations in the temperature must also be considered as a quantum operator. The observables are often expressed as
ˆ
n-point correlation functions of the operator ∆(e) ≡ δTˆ/T (e) in the arbitrary state |Ψi

ˆ 1 ) · · · ∆(e
ξn (e1 , · · · , en ) ≡ hΨ|∆(e ˆ n )|Ψi, (1)

where ei ’s are arbitrary directions on the celestial sphere. In the following, we will also use the notation ξ(e1 · e2 ) ≡
ξ2 (e1 , e2 ). According to the postulates of quantum mechanics, the previous theoretical predictions should be confronted
to experiment in the following way. The same experiment should be performed N times giving each time different
PN
outcomes qi . If the quantity (1/N ) i=1 qi goes to the corresponding quantum expectation value when N goes to infinity,
the theoretical prediction is said to be ‘compatible with experiment’. This is the core of the problem in cosmology: we only
4

have access to one realization, i.e. one map of the CMB sky and that means that N is fixed and equal to one. Therefore,
the question arises as to how we can verify the theoretical predictions of the theory of quantum-mechanical cosmological
perturbations. This is a way of stating the cosmic variance problem. It is a fundamental limitation in the sense that it
remains even when other limitations like instrumental errors or low angular resolution have been fully mastered.
The usual method to deal with this problem is to replace quantum averages with spatial averages over the celestial
sphere. Suppose we wish to measure ξn (e1 , · · · en ). (Of course, the discussion could also be applied to quantities other
than correlation functions). The first step is to introduce a new operator, the estimator Ê(ξn ) of ξn , defined as
Z Z
ˆ ˆ 1 ) · · · ∆(e
E(ξn ) ≡ · · · dΩ1 · · · dΩn E(ξn )(e1 · · · en )∆(e ˆ n ), (2)

where E(ξn )(e1 · · · en ) is a weight function to be determined. Clearly, Ê(ξn ) is defined through a spatial average. The
second step is to require that the estimator is unbiased, i.e.

hΨ|Ê(ξn )|Ψi = ξn . (3)

In general, this restricts the class of functions E(ξn ) allowed. The fact that the mean value of the estimator be equal to
the quantity we are seeking does not guarantee that each outcome will be for sure ξn . The third and final step is then to
ˆ n ), i.e.
find the function E(ξn ) such that the variance (squared) of E(ξ

2


2
σÊ(ξ = Ê(ξn )Ê(ξn ) − Ê(ξn ) (4)
n)

be as small as possible, taking into account the constraint given by Eq. (3). The corresponding estimator is then called
the best unbiased estimator. Mathematically, this requirement is expressed through the following variation equation
  
2
δ σÊ(ξ )
− λ hΨ| Ê(ξn )|Ψi − ξn = 0, (5)
n

where one has introduced a Lagrange multiplier λ which can then be determined from the previous equation and the
constraint itself. Once we have λ, we plug it into Eq. (5) and this completely fixes E(ξn ) and hence the corresponding
best estimator. In turn, its variance can now be calculated. If this one vanishes then we are sure that each outcome is
ˆ
ξn and from one realization we can determine the n-point correlation function. In this case, ∆(e) is said to be ergodic,
i.e. ensemble or quantum averages coincide with spatial averages. Unfortunately, one can show that this cannot be the
case on the two-dimensional sphere (Grishchuk & Martin 1997). Otherwise, we have found the weight function E(ξn )
which leads to the smallest non-vanishing variance. If the variance is small enough, each outcome will be concentrated
around the mean value and with just one realization we have good chances to get a reasonable estimate of the correlation
function. The typical error made in considering that one given outcome is equal to the mean value is characterized by
the variance of the estimator, see Fig. 1.
All the above analysis performed for the best (quantum) estimators can equally well be reproduced in the case where
the anisotropies are due to an underlying stochastic process although the former is generally physically best motivated.
In that case, quantum averages hΨ| · · · |Ψi are just replaced with stochastic averages h· · ·i. In the following, we will drop
out the ‘hat’ symbol and consider that the different quantities are either operators or stochastic processes. In the same
manner, we will denote an ensemble average by the symbol h· · ·i having in mind that this means either quantum or
classical averages.
Let us now describe the relevant quantities to estimate. It is convenient to expand the temperature fluctuations over
the basis of spherical harmonics according to
X m
∆(e) = am
ℓ Yℓ (e). (6)
ℓm
 5

σε
Best( ξn )

σε ( ξn )

ξn
Fig. 1.— Sketch comparing the variances of an arbitrary estimator of ξn with the best estimator of the same quantity. The widest
distribution does not permit an accurate determination of ξn due to its large variance whilst the narrow distribution corresponds to the best
(unbiased) estimator and possesses the smallest possible variance. As a consequence, one given realization (i.e. our CMB sky) will most
probably be closest to the mean value in the second case rather than in the first one.

This equation assumes a complete sky coverage. Implementation of the method for the incomplete (galaxy-cut) sky can
be performed by using the basis introduced in (Górski 1994). Once a specific model is given, the statistical properties of
the am m m∗
ℓ ’s are determined. Since ∆(e) is real, the aℓ ’s must satisfy aℓ = (−1)m a−m
ℓ . Without restricting the generality
of the underlying physics, the first three moments can be written as

m
m1 m2 ∗ 
m1 m2 m3  ℓ1 ℓ2 ℓ3


aℓ = 0, aℓ 1 aℓ 2 = C ℓ 1 δℓ 1 ℓ 2 δm1 m2 , aℓ 1 aℓ 2 aℓ 3 = m1 m2 m3 C ℓ1 ℓ2 ℓ3 , (7)

is a Wigner 3j-symbol. The second equation ensures the isotropy of the CMB. The quantity am
ℓ1 ℓ2 ℓ3


1 m2 ∗ 
where m1 m2 m3 ℓ 1 aℓ 2
is the second moment of the amℓ ’s and Cℓ is usually called the angular spectrum. In the third equation, the quantity

m1 m2 m3 
aℓ1 aℓ2 aℓ3 is the third moment while Cℓ1 ℓ2 ℓ3 is called the angular bispectrum. For ℓ1 = ℓ2 = ℓ3 = ℓ, this quantity is
generally written as Bℓ ≡ Cℓℓℓ . The presence of the Wigner 3j-symbol guarantees that the third moment vanishes unless
m1 + m2 + m3 = 0 and |ℓi − ℓj | ≤ ℓk ≤ ℓi + ℓj . Moreover, invariance under spatial inversions of ξ3 implies an additional
‘selection rule’ (Luo 1994, Gangui & Martin 2000), ℓ1 + ℓ2 + ℓ3 = even, in order for the third moment not to vanish.
Finally, from this last relation and using standard properties of the 3j-symbols, it follows that the angular bispectrum is
left unchanged under any arbitrary permutation of the indices ℓi .
We will need the higher moments as well. Since departures from Gaussianity are expected to be small (specially on large
angular scales), higher moments will be calculated in the mildly non-Gaussian approximation. Within this approximation
m(0) m(1) m(0)
we can write am
ℓ = aℓ + ǫ aℓ + O(ǫ2 ) where aℓ is a Gaussian random variable and the expansion parameter ǫ
is small. In the following, each moment will be calculated to the first non-vanishing order in ǫ. For example, the fourth

1 m2 m3 m4 
m1 (0) m2 (0) m3 (0) m4 (0) 
moment yields am ℓ 1 aℓ 2 aℓ 3 aℓ 4 = aℓ 1 aℓ 2 aℓ 3 aℓ 4 + O(ǫ). As a consequence, the connected fourth moment
can be neglected because it is of higher order than the Gaussian part. The ‘(0) ’ label will be dropped out hereafter.
Therefore, in the mildly non-Gaussian approximation we can write

am 1 m2 m3 ∗ m4 ∗


m1 m2 
m3 ∗ m4 ∗ 
m1 m3 ∗ 
m2 m4 ∗ 
m1 m4 ∗ 
m2 m3 ∗ 
ℓ 1 aℓ 2 aℓ 3 aℓ 4 ≈ aℓ 1 aℓ 2 aℓ 3 aℓ 4 + aℓ 1 aℓ 3 aℓ 2 aℓ 4 + aℓ 1 aℓ 4 aℓ 2 aℓ 3
= (−1)m2 +m4 Cℓ1 Cℓ3 δℓ1 ℓ2 δm1 ,−m2 δℓ3 ℓ4 δm3 ,−m4 + Cℓ1 Cℓ2 δℓ1 ℓ3 δm1 m3 δℓ2 ℓ4 δm2 m4
+ C ℓ 1 C ℓ 2 δℓ 1 ℓ 4 δm1 m4 δℓ 2 ℓ 3 δm2 m3 . (8)

The fifth moment could be determined in a similar way but we will not need this quantity in the following. Finally, the
6

sixth moment can be expressed as

m1 m2 m3 m4 ∗ m5 ∗ m6 ∗ 
m1 m2 
m3 m4 ∗ 
m5 ∗ m6 ∗ 
aℓ 1 aℓ 2 aℓ 3 aℓ 4 aℓ 5 aℓ 6 ≈ aℓ 1 aℓ 2 aℓ 3 aℓ 4 aℓ 5 aℓ 6 + 14 additional permutations . (9)

Although the explicit expression is not particularly illuminating, the last equation will be useful for the calculation of the
variance when dealing with the three-point correlators below. In particular, one can write

am 1 m2 m3 m1 ∗ m2 ∗ m3 ∗
= Cℓ1 Cℓ2 Cℓ3 + 2Cℓ31 δℓ1 ℓ2 ℓ3 (δm1 m3 + δm1 −m3 )(δm1 m2 + δm1 −m2 )


ℓ 1 aℓ 2 aℓ 3 aℓ 1 aℓ 2 aℓ 3

+ Cℓ1 Cℓ22 δℓ2 ℓ3 (δm2 m3 + δm2 −m3 ) + Cℓ2 Cℓ23 δℓ3 ℓ1 (δm1 m3 + δm1 −m3 )
+ Cℓ3 Cℓ21 δℓ1 ℓ2 (δm1 m2 + δm1 −m2 ), (10)

where the symbol δℓ1 ℓ2 ℓ3 vanishes unless ℓ1 = ℓ2 = ℓ3 in which case it is one. This equation coincides with Eq. (24)
of (Luo 1994) provided the undefined symbol δm1 m2 m3 ,0 written in that work has the meaning δm1 m2 m3 ,0 ≡ (δm1 m3 +
δm1 −m3 )(δm1 m2 + δm1 −m2 ).
As we mentioned in the Introduction, in this article we are mainly interested in finding the best unbiased estimators for
the two following quantities: the third moment am

1 m2 m3 
ℓ 1 aℓ 2 aℓ 3 and the angular bispectrum Cℓ1 ℓ2 ℓ3 . These are related to
the three-point correlation function ξ3 . It is clear that, as mentioned above, the very same method could also be utilized
for the computation of ξn with n arbitrary. Before addressing this question, however, we will first treat the analogous
quantities related to the two-point correlation function, namely am

1 m2 ∗ 
ℓ 1 aℓ 2 and Cℓ , the main purpose being to illustrate
concretely the tactics presented above in a case where everything can be calculated easily. This will be used as a guideline
for the case of the three-point correlators.

3. TWO-POINT CORRELATORS

3.1. Best estimator for the angular spectrum Cℓ

The definition of the estimator E(Cℓ ) is given by our general prescription, see Eq. (2)
Z Z
E(Cℓ ) ≡ dΩ1 dΩ2 E ℓ (e1 , e2 )∆(e1 )∆(e2 ), (11)

where E ℓ (e1 , e2 ) is the weight function. The angular spectrum Cℓ is a real quantity and its estimator E(Cℓ ) must also be
real. Therefore, the weight function can be taken real. From the previous definition, it is clear that the antisymmetric
part of E ℓ (e1 , e2 ) does not contribute to the estimator. Then, we can replace E ℓ (e1 , e2 ) in E(Cℓ ) by its symmetrized
expression ESℓ (e1 , e2 ) = (1/2)[E ℓ (e1 , e2 ) + E ℓ (e2 , e1 )]. At this stage, two methods can be applied. Either we work directly
with the weight function or we expand it over the spherical harmonics basis and try to determine the coefficients of this
expansion. Clearly, both paths are equivalent and can be followed for any estimator. Here we employ the second method,
leaving the first one for the determination of the second-moment estimator considered in the next subsection. Therefore,
we write the weight function as
X X
ESℓ (e1 , e2 ) = dnℓ ℓ o Y m1 (e1 )Y m2 (e2 ). (12)
1 ℓ2 ℓ1 ℓ2
ℓ 1 m1 ℓ 2 m2 m1 m2

The reality and symmetry properties of the weight function ESℓ imply that the complex coefficient of the expansion must
satisfy, respectively
dnℓ∗ℓ ℓ2
o = (−1)m1 +m2 dnℓ
ℓ1 ℓ2
o , dnℓ ℓ ℓ2
o = dnℓ
ℓ ℓ1
o. (13)
1 1 2
m1 m2 −m1 −m2 m1 m2 m2 m1
 7

Inserting the expression of the weight function (12) into the general definition of the estimator (11) and using standard
properties of the spherical harmonics, one gets
X X
E(Cℓ ) = dnℓ∗ℓ o am1 am2 . (14)
1 ℓ2 ℓ1 ℓ2
ℓ 1 m1 ℓ 2 m2 m1 m2

Our first move is now to require that hE(Cℓ )i = Cℓ . Using the second of Eqns. (7), we find that the coefficients d must
fulfill the following constraints
X
(−1)m1 dnℓ ℓ1 ℓ1
o = δℓℓ .
1
(15)
m1 −m1 m1

All the estimators satisfying this condition are unbiased. However, it is clear that this does not completely determines the
estimator but just a class of estimators. Our second move is to calculate the variance. Using Eq. (8), a straightforward
computation gives
X X
2
σE(Cℓ)
=2 dnℓ∗ℓ ℓ2
o dnℓ
ℓ1 ℓ2
o C ℓ1 C ℓ2 . (16)
1
ℓ 1 m1 ℓ 2 m2 m1 m2 m1 m2

This quantity is obviously positive. From this expression, one sees that the imaginary part of the coefficients d only
increases the variance. Since a vanishing ℑ(d) satisfies the constraint equation, we can consider that the d’s are real.
Our third move is to minimize the variance taking into account the constraint. For this purpose we introduce a set of
Lagrange multipliers λℓℓ1 (i.e. one Lagrange multiplier per constraint since ℓ must be seen as a fixed index) and require
that  X 
X
2 ℓ
δ σE(Cℓ)
+ λℓ1 (−1)m1 dnℓ ℓ1 ℓ1
o − δℓℓ
1
= 0. (17)
ℓ1 m1 −m1 m1

The definition of the variation δ must respect the symmetry properties of the coefficients d; we take

δdnℓ ℓ o
1 ℓ2   
m1 m2 1 1 m 1 +m2
≡ δℓ ℓ′ δℓ ℓ′ δm1 m′1 δm2 m′2 + (−1) δm1 −m′1 δm2 −m′2
δdnℓ ℓ′ ℓ′
o 2 2 11 22
1 2
m′ m′
1 2
 
1
+ δℓ1 ℓ′2 δℓ2 ℓ′1 δm1 m′2 δm2 m′1 + (−1)m1 +m2 δm1 −m′2 δm2 −m′1 . (18)
2
Although it is not compulsory to use this equation, since a naive definition of the variation would lead to the same final
result (Grishchuk & Martin 1997), it is nevertheless interesting to utilize it as a warm up for what will be done for the
angular bispectrum. The variation leads to the following relation between the coefficients d and the Lagrange multipliers

4dnℓ ℓ ℓ2
o Cℓ1 Cℓ2 + (−1)m1 λℓℓ δℓ1 ℓ2 δm1 −m2 = 0.
1
(19)
1
m1 m2

We see that the choice of the λℓ ’s is not free; it is fixed by the variation itself. Using the constraint in this equation,
we find λℓℓ1 = −4[Cℓ21 /(2ℓ1 + 1)]δℓℓ1 . Having determined what the Lagrange multipliers are, the problem is completely
solved. It is now sufficient to re-introduce this value for λℓℓ1 in Eq. (19), get the coefficients d and, from this, also the
ℓ
best unbiased estimator: the weight function is then given by ES,Best (e1 , e2 ) = (1/4π)Pℓ (e1 · e2 ) and the estimator itself
by (see also Grishchuk & Martin 1997)
ℓ
1 X
EBest (Cℓ ) = am m∗
ℓ aℓ . (20)
2ℓ + 1
m=−ℓ

The variance of this estimator is the well-known ‘cosmic variance’

2Cℓ2
σE2Best (Cℓ ) = . (21)
2ℓ + 1
8

One remark is in order here. The cosmic variance is usually obtained in the following way: the previous estimator appears
naturally from Eqns. (7) and it is usually assumed that the am
ℓ ’s are Gaussian random variables. In this case, the estimator
(20) has a χ2 probability density function and from this the cosmic variance can be easily recovered. The proof presented
above is by no means equivalent to this naive derivation. There are many unbiased estimators and, a priori, nothing
guarantees that the simplest one is the best one, i.e. the naive derivation is not sufficient to prove that the estimator
(20) is the best one. This can be proven only along the lines described above. Moreover, we do not need to assume that
the am
ℓ ’s obey a Gaussian statistics. Only the mildly non-Gaussian assumption is necessary for the calculation of the
four-point correlators.

3.2. Best estimator for the second moment Cα

In this section, we discuss the best estimator for Cα ≡ am

1 m2 ∗ 
n o n ′ ′o ℓ 1 aℓ 2 . Greek letters will always be employed for collective-
ℓ1 ℓ2 ′ ℓ ℓ
index notation, like α ≡ m1 m2 or α ≡ m′ m′ . Unlike in the foregoing subsection, we find the weight function
1 2
1 2

directly without expanding it over the spherical harmonics basis. This method is closer to the one used by Heavens
(1998). We also show that the best estimator of Cα cannot be deduced from the best estimator of the angular spectrum
Cℓ obtained in the previous subsection as one could naively think.
Let us start with a couple of definitions. First, the quantity

Rα (e1 , e2 ) ≡ Yℓm
1
1
(e1 )Yℓm
2
2∗
(e2 ), (22)

which is complex and non-symmetric with regards to the position of the complex conjugate symbol ‘∗ ’. This last property
comes from the definition of am

1 m2 ∗ 
ℓ 1 aℓ 2 itself and it would be the case for any even-point correlator. Then, regarding
complex conjugation, there is a slight difference between the even- and the odd-point correlators. The real part of
Rα (e1 , e2 ) will be noted  
1 m1
α
RR (e1 , e2 ) ≡ Yℓ1 (e1 )Yℓm 2∗
(e 2 ) + Y m1 ∗
ℓ1 (e )Y
1 ℓ2
m2
(e 2 .
) (23)
2 2

α α ᾱ
RR (e1 , e2 ) is not symmetric under a permutation of the two directions. It satisfies RR (e2 , e1 ) = RR (e1 , e2 ), where the
ℓ2 ℓ1
index ᾱ is defined by ᾱ ≡ { m2 m1
}. The symmetries in the indices and in the directions will play an important rôle in
what follows. Using the previous definitions, we can introduce a quantity which is symmetric both under a permutation
of the two directions and under a permutation of the columns of internal indices in α
 
1 m1
RSα (e1 , e2 ) ≡ Yℓ1 (e1 )Yℓm 2∗
(e 2 ) + Y m1
ℓ1 (e )Y
2 ℓ2
m2 ∗
(e 1 ) + Y m1 ∗
ℓ1 (e )Y
1 ℓ2
m2
(e 2 ) + Y m1 ∗
ℓ1 (e )Y
2 ℓ2
m2
(e 1 ,
) (24)
4 2

i.e. we have RSα (e1 , e2 ) = RSα (e2 , e1 ) = RSᾱ (e1 , e2 ) = RSᾱ (e2 , e1 ). Finally, we will also employ a symmetrized Krönecker
symbol defined according to
 
′ 1
δSαα ≡ δℓ1 ℓ′1 δℓ2 ℓ′2 δm1 m′1 δm2 m′2 + δℓ1 ℓ′2 δℓ2 ℓ′1 δm1 m′2 δm2 m′1 , (25)
2

which is left unchanged under a permutation of the indices α and α′ and also under permutations of the columns of each
collective index separately.
Let us now turn to the computation of the best unbiased estimator; the general definition of an estimator can be
expressed as Z Z
E(Cα ) ≡ dΩ1 dΩ2 E α (e1 , e2 )∆(e1 )∆(e2 ). (26)

The quantity Cα is unchanged if we permute the columns of indices in α, Cα = Cᾱ with ᾱ given above. Being an estimator
for Cα it is then natural to assume that E(Cα ) possesses the same property, E(Cα ) = E(Cᾱ ). Looking at the definition of
 9

E α (e1 , e2 ), Eq. (26), we easily see that the weight function will also satisfy E α (e1 , e2 ) = E ᾱ (e1 , e2 ). Moreover, we take
E α (e1 , e2 ) symmetric under a permutation in the directions e1 and e2 , i.e. ESα (e1 , e2 ) = ESα (e2 , e1 ).
Now, we require that the estimator E(Cα ) be unbiased, which implies that the following relation must be fulfilled
Z Z
′ ′
dΩ1 dΩ2 ESα (e1 , e2 )Rα (e1 , e2 ) = δSαα . (27)

′
We have required the presence of δSαα in the right hand side of the previous equation to respect the symmetries in α of
′
the weight function. In the previous equation the ESα and Rα are a priori complex. However, one can show that it is
possible to work only with a real weight function, ESα∗ (e1 , e2 ) = ESα (e1 , e2 ) and with the RR
α
(e1 , e2 ) defined above: the
imaginary contributions would just increase the variance. (One could have also chosen to work with a pure imaginary
weight function and the imaginary part of Rα ). Therefore, the constraint can be written as
Z Z
α′ ′
dΩ1 dΩ2 ESα (e1 , e2 )RR (e1 , e2 ) = δSαα , (28)

where ESα has been taken real. Let us now calculate the variance of the estimator E(Cα ). Using Eqns. (8), we easily find
that Z Z Z Z
2
σE(Cα)
=2 dΩ1 dΩ2 dΩ3 dΩ4 ESα (e1 , e2 )ESα (e3 , e4 )ξ(e1 ·e3 )ξ(e2 ·e4 ). (29)

Our next step is to minimize this variance under the constraint given in Eq. (28). For this purpose, we introduce a set of
Lagrange multipliers λα and require that
 X Z Z 
2 α α α′ αα′
δ σE(Cα)
+ λα′ dΩ 1 dΩ E
2 S (e ,
1 2e )RR (e ,
1 2e ) − δ S = 0. (30)
α′

At this point the precise meaning of the variation symbol δ matters. Before performing the variation, let us recall that
the symmetries of the weight function must be respected; hence we will have
δESα (ei , ej )
 
1 δ(ei ·ek − 1) δ(ej ·eℓ − 1) δ(ei ·eℓ − 1) δ(ej ·ek − 1) αβ
= + δS . (31)
δESβ (ek , eℓ ) 2 2π 2π 2π 2π

The 2π’s in the denominators come from the fact that, while the direction ei is expressed in terms of the corresponding
spherical angles like ei ≡ (θi , ϕi ) and then dΩi = d cos θi dϕi , there is an extra 2π factor in δ(ei ·ek − 1) = 2π δ(cos θi −
cos θk )δ(ϕi − ϕk ). As a result of the variation, we obtain the following equation
1 X α α′
Z Z
dΩ1 dΩ2 ESα (e1 , e2 )ξ(e1 ·e3 )ξ(e2 ·e4 ) = λ ′ R (e3 , e4 ). (32)
4 ′ α S
α
′
The quantity RSα (e3 , e4 ) appears naturally as a result of Eq. (31). The previous equation should be compared with Eq.
(19) of the previous subsection. The result of the variation is a relation between the weight function and the Lagrange
multiplier. Our aim now is to get an explicit expression for the weight function. This can be done by using the inverse
two-point correlation function ξ −1 which satisfies (see also the Appendix)
Z
dΩj ξ(ei ·ej )ξ −1 (ej ·ek ) ≡ δ(ei ·ek − 1). (33)

In the case of the three-point correlator, this definition leads to subtleties which will be examined in detail in the next
section. Multiplying Eq. (32) by ξ −1 (e1′ · e3 )ξ −1 (e2′ · e4 ), integrating over directions e3 and e4 and relabelling indices,
we arrive at
1 1 X
Z Z
′
ESα (e1 , e2 ) = 2
dΩ3 dΩ4 λα
α′ ξ
−1
(e1 ·e3 )ξ −1 (e2 ·e4 )RSα (e3 , e4 ). (34)
4 (2π) ′α

This equation for the second moment is the analogous of Eq. (21) of (Heavens 1998) obtained for the third moment. It
is clear from the previous section that, at this stage, our final goal has not yet been reached. The weight function is still
10

expressed in terms of the Lagrange multipliers. The correct way to proceed is to remove the latter using the constraint
′′
given by Eq. (28) as it was done in the previous subsection. Then, we first multiply Eq. (34) by the quantity RSα (e1 , e2 )
and, after having integrated over solid angles Ω1 and Ω2 , this leads to the equation that the Lagrange multipliers must
satisfy
 
′′
m′′ ′′ ′′ ′′
16Cℓ′′1 Cℓ′′2 δSαα = (−1) 1 +m2 λα
n
ℓ′′ ℓ′′
o + (−1)m1 +m2 λα
n
ℓ′′ ℓ′′
α
o + λn
ℓ′′ ℓ′′
o + λα
n
ℓ′′ ℓ′′
o . (35)
1 2 2 1 1 2 2 1
−m′′ −m′′ −m′′ −m′′ m′′ m′′ m′′ m′′
1 2 2 1 1 2 2 1

This equation is the analogue of the equation λℓℓ1 = −4[Cℓ21 /(2ℓ1 + 1)]δℓℓ1 of the previous subsection. Here, the difference
is that a combination of Lagrange multipliers with different indices appears rather than the Lagrange multiplier itself.
However, we can reconstruct exactly this combination in the right hand side of Eq. (34). Indeed, we just have to multiply
′′ ′′
each side of Eq. (35) by RSα (e1 , e2 ) and perform the sum over α′′ . Using the symmetry properties of RSα (e1 , e2 ) we
obtain
X ′′
4Cℓ1 Cℓ2 RSα (e1 , e2 ) = λα α
α′′ RS (e1 , e2 ). (36)
α′′

We now insert this equation in the right hand side of Eq. (34) to remove the Lagrange multipliers and find
1
Z Z
α
ES,Best (e1 , e2 ) = Cℓ Cℓ dΩ3 dΩ4 ξ −1 (e1 ·e3 )ξ −1 (e2 ·e4 )RSα (e3 , e4 ). (37)
4π 2 1 2
To go further, we express ξ −1 in Eq. (37) explicitly in terms of spherical harmonics (see the Appendix) which yields
α
ES,Best (e1 , e2 ) = RSα (e1 , e2 ). Now, we just plug this result into Eq. (26) to get the final explicit expression of the best
unbiased estimator EBest (Cα )  
1 m1 ∗ m2
EBest (Cα ) = aℓ 1 aℓ 2 + am 1 m2 ∗
a
ℓ1 ℓ2 . (38)
2
Some remarks are in order here. First, despite appearances, one cannot deduce EBest (Cα ) from EBest (Cℓ1 ). It is clear that
the following equation holds
hE(Cα )i = hE(Cℓ1 )iδℓ1 ℓ2 δm1 m2 . (39)

However, from this equation we are not allowed to conclude that EBest (Cα ) = EBest (Cℓ1 )δℓ1 ℓ2 δm1 m2 . Therefore, knowing
one of the best unbiased estimators does not allow us to infer the other best one. To be specific, if one assumed the
previous wrong relation, from Eq. (20) one would get
ℓ1
 X 
1
EBest (Cα ) = am∗ m
ℓ 1 aℓ 1 δℓ 1 ℓ 2 δm1 m2 (false), (40)
2ℓ1 + 1
m=−ℓ1

and although this estimator is unbiased, it is not the best one. The second remark is that although the estimator given by
Eq. (38) could be naively regarded as a trivial one, this is not the case. Indeed, the estimator of Eq. (40) is as trivial as
the actual best one. The moral is then that there exist simple choices which lead to the wrong answer. The only reliable
method in the problem of minimizing the variance is therefore the one exposed above.

4. THREE-POINT CORRELATORS

4.1. Best estimator for the angular bispectrum Cℓ1 ℓ2 ℓ3

In this section, our aim is to determine the best unbiased estimator for the angular bispectrum Cℓ1 ℓ2 ℓ3 defined in the
third of Eqns. (7). According to our general prescription, the most general definition reads
Z Z Z
E(Cℓ1 ℓ2 ℓ3 ) ≡ dΩ1 dΩ2 dΩ3 ESℓ1 ℓ2 ℓ3 (e1 , e2 , e3 )∆(e1 )∆(e2 )∆(e3 ). (41)
 11

As in the case of the two-point correlators, the weight function also possesses the properties of being real and symmetric
under arbitrary permutations of directions ei . In addition, like Cℓ1 ℓ2 ℓ3 , the weight function satisfies ESℓ1 ℓ2 ℓ3 = ESℓ2 ℓ1 ℓ3 , as
well as for any other arbitrary permutation of the indices ℓi . We follow similar steps as for the angular spectrum and
therefore we choose to expand the weight function on the basis of the spherical harmonics. Then, as in Eq. (12), we write
′ ′ ′
o Yℓm m m
X X X
ESℓ1 ℓ2 ℓ3 (e1 , e2 , e3 ) =
ℓ ℓ2 ℓ3
dn1ℓ′ ℓ′ ℓ′
′
1
(e1 )Yℓ′ 2 (e2 )Yℓ′ 3 (e3 ). (42)
1 2 3 1 2 3
ℓ′1 m′1 ℓ′2 m′2 ℓ′3 m′3 m′
1
m′
2
m′
3

The properties of the weight function imply that the coefficients d must satisfy equations similar to those given in Eqns.
(13)
ℓ ℓ2 ℓ3 ∗ ′ ′ ′
dn1ℓ′ o = (−1)m1 +m2 +m3 dnℓ1 ′ ℓ2 ℓ3
o ,
ℓ
dn1ℓ′
ℓ2 ℓ3
o = dnℓ1 ′ ℓ2 ℓ3
o, (43)
ℓ′ ℓ′ ℓ ℓ′ ℓ′ ℓ′ ℓ′ ℓ ℓ′ ℓ′
1 2 3 1 2 3 1 2 3 2 1 3
m′ m′ m′ −m′ −m′ −m′ m′ m′ m′ m′ m′ m′
1 2 3 1 2 3 1 2 3 2 1 3

where the last relation is in fact valid for arbitrary permutations of any two columns of the collective subindex. Like the
weight function, d is also left invariant under arbitrary permutations of indices ℓi (not primed). The estimator can be
expressed in terms of the coefficients d and the am
ℓ ’s only: inserting the expansion of the weight function in the above
expression for the estimator and using standard properties of the spherical harmonics one obtains
′ ′ ′
o am 1 m2 m3
X X X ℓ ℓ2 ℓ3 ∗
E(Cℓ1 ℓ2 ℓ3 ) = dn1ℓ′ ℓ′ ℓ′ ℓ ′ aℓ ′ aℓ ′ . (44)
1 2 3 1 2 3
ℓ′1 m′1 ℓ′2 m′2 ℓ′3 m′3 m′
1
m′
2
m′
3

In practice, CMB observational settings are devised such that both the monopole and the dipole are subtracted from
the anisotropy maps. This means that the coefficients d in the last equation are only non-vanishing for indices ℓ′i ≥ 2 in
the collective subindex. Moreover, the coefficients d satisfy ℓ1 + ℓ2 + ℓ3 = even. We must now require that our general
estimator given by Eq. (44) be unbiased, i.e. hE(Cℓ1 ℓ2 ℓ3 )i = Cℓ1 ℓ2 ℓ3 . This forces the coefficients d to fulfill the following
constraint  
ℓ′1 ℓ′2 ℓ′3 ′
 = δ ℓi ℓj ,
X ℓ1 ℓ2 ℓ3 ∗
dn ℓ′ ℓ′ ℓ′
o S (45)
m′1 m′2 m′3
1
m′
2
m′
3
m′
m′1 m′2 m′3
1 2 3

where we have defined a new symmetrized Krönecker symbol, this time for the ℓ multipole indices only, as follows
 
ℓi ℓ′ 1
δS j ≡ δℓ1 ℓ′1 δℓ2 ℓ′2 δℓ3 ℓ′3 + 5 additional permutations . (46)
6

It is easy to check that the constraint equation satisfies the conditions imposed by Eqns. (43) on the coefficients d. In
ℓi ℓ′j
particular, let us justify the presence of the symbol δS . Using the previous properties for d, relabelling the indices
m′1 ↔ m′2 in Eq. (45) and finally noting that ℓ′1 + ℓ′2 + ℓ′3 = ℓ1 + ℓ2 + ℓ3 = even, which allows us to permute any two
columns of the Wigner 3j-symbol, one verifies that the left hand side of the constraint is invariant under ℓ′1 ↔ ℓ′2 . The
ℓi ℓ′j
same applies for any pair of ℓ multipole indices and this explains the presence of the symmetrized δS in Eq. (45). We
see from this that all coefficients d that do not satisfy ℓ′1 + ℓ′2 + ℓ′3 = even do not enter the constraint. We will show below
that these terms only increase the variance and as a consequence one can take them equal to zero. In particular, note
that this is the case for a coefficient d with ℓ′1 = ℓ′2 and ℓ′3 = odd. This property will turn out to be useful in what follows.
We are now in a position to calculate the variance of the estimator. Looking at Eq. (44) we see that this requires the
computation of the sixth moment of the am
ℓ ’s, see Eq. (9). After having made use of the properties of the coefficients d
and rearranging the resulting 15 terms into two groups, straightforward algebra yields
 
2
X X X ℓ ℓ ℓ ∗ ℓ ℓ ℓ ′ ′ ℓ ℓ2 ℓ3 ∗
h[E(Cℓ1 ℓ2 ℓ3 )] i = Cℓ1 Cℓ2 Cℓ3 6dn1ℓ′ 2ℓ′ 3ℓ′ o dn1ℓ′ 2ℓ′ 3ℓ′ o + 9(−1)m1 +m2 dn1ℓ′
′ ′ ′ o dnℓ1 ′ ℓ2 ℓ3
o .
ℓ′ ℓ′ ℓ ℓ′ ℓ′
1 2 3 1 2 3 1 1 3 2 2 3
ℓ′1 m′1 ℓ′2 m′2 ℓ′3 m′3 m′
1
m′
2
m′
3
m′
1
m′
2
m′
3
m′
1
−m′
1
m′
3
m′
2
−m′
2
m′
3
(47)
12

The square of the variance of E(Cℓ1 ℓ2 ℓ3 ) is given by

2

2
2
σE(Cℓ ) = [E(Cℓ1 ℓ2 ℓ3 )] − E(Cℓ1 ℓ2 ℓ3 ) . (48)
1 ℓ2 ℓ3

2 2
Following the discussion in §2, the term [E(Cℓ1 ℓ2 ℓ3 )] is of order ǫ0 whereas the lowest non-vanishing order of E(Cℓ1 ℓ2 ℓ3 )

is ǫ2 . Therefore, the latter one will not enter the minimization procedure and the variance squared will be written as
2

2
σE(C ℓ ℓ ℓ )
≈ [E(Cℓ1 ℓ2 ℓ3 )] . However, let us notice that this does not occur in the case of the two-point correlator. Indeed,
1 2 3

as we have seen, in that case both terms contributing to the square of the variance are of the same order in ǫ. Then, it
follows that Eq. (47) corresponds to Eq. (16) in the last section.
Let us now examine the structure of the variance in more detail. The term 6d∗ d ∝ ℜ2 (d) + ℑ2 (d) in Eq. (47) is
analogous to the one in Eq. (16). However, here there is another contribution, the 9d∗ d term, which will play a crucial
rôle in what follows. The imaginary part of this term of course vanishes, as the variance must be real. However, the real
part of it contains a contribution like
     
[. . .]2 ,
X X X X ′ X X
m′2
ℓ ℓ2 ℓ3 ℓ ℓ2 ℓ3
X X
9 Cℓ′3  Cℓ′1 (−1)m1 ℑ dn1ℓ′ ℓ′ ℓ′
o   Cℓ′2 (−1) ℑ dn1ℓ′ ℓ′ ℓ′
o  = 9 Cℓ′3
1 1 3 2 2 3
ℓ′3 m′3 ℓ′1 m′1 m′
1
−m′
1
m′
3
ℓ′2 m′2 m′
2
−m′
2
m′
3
ℓ′3 m′3
(49)
2
where the quantity [. . .] depends on the indices ℓ1 , ℓ2 , ℓ3 and ℓ′3 , m′3 and is strictly positive or zero. Of course, there is
a similar term coming from the real part of d. Thus, we see that the various contributions of the imaginary part of the
coefficients d to the two terms, 6d∗ d and 9d∗ d, only increase the variance. Since we know that a vanishing imaginary part
does satisfy the constraint Eq. (45), it can be disregarded in the sequel. Therefore, Eq. (47) can then be written solely
in terms of real coefficients d as follows
X X X   2 
ℓ ℓ2 ℓ3 ′ ′ ℓ ℓ2 ℓ3
2
σE(C ) = Cℓ′1 Cℓ′2 Cℓ′3 6 dn1ℓ′ o + 9(−1)m1 +m2 dn1ℓ′ o dnℓ1 ′ ℓ2 ℓ3
o . (50)
ℓ 1 ℓ2 ℓ3 ℓ′ ℓ′ ℓ′ ℓ′ ℓ ℓ′ ℓ′
1 2 3 1 1 3 2 2 3
ℓ′1 m′1 ℓ′2 m′2 ℓ′3 m′3 m′
1
m′
2
m′
3
m′
1
−m′
1
m′
3
m′
2
−m′
2
m′
3

Our next move now is to minimize this variance with respect to the coefficients d, taking into account the constraint of
Eq. (45)  
ℓ′1 ℓ′2 ℓ′3
  X ′

 − δ ℓi ℓj
X
λℓℓ′1 ℓℓ′2 ℓℓ′3
2 ℓ1 ℓ2 ℓ3
δ σE(Cℓ ) + dn ℓ′ ℓ′ ℓ′
o S = 0. (51)
1 ℓ2 ℓ3
ℓ′1 ℓ′2 ℓ′3
1 2 3
m′1 m′2 m′3
1
m′
2
m′
3
m′
m′1 m′2 m′3
1 2 3

This equation is the analogous to Eq. (17). As before, one needs to give a concrete meaning to the symbol δ of the
variation. Its definition must respect the symmetries of the coefficients d and hence we take
ℓ ℓ2 ℓ3
δdn1ℓ′ ℓ′ ℓ′
o
1 2 3    
m′
1
m′
2
m′
3
1 1 m ′
+m ′
+m ′
= δℓ′ ℓ′′ δℓ′ ℓ′′ δℓ′ ℓ′′ δm′1 m′′1 δm′2 m′′2 δm′3 m′′3 + (−1) 1 2 3 δm′1 −m′′1 δm′2 −m′′2 δm′3 −m′′3 + 5 terms . (52)
ℓ ℓ ℓ
δdn1ℓ′′ 2 ℓ′′ 3 ℓ′′ o 6 2 11 22 33
1 2 3
m′′ m′′ m′′
1 2 3

Then, Eq. (51) leads to

 
ℓ′1 ℓ′2 ℓ′3 X
12Cℓ′1 Cℓ′2 Cℓ′3 dn1ℓ′2ℓ′3ℓ′ o + λℓℓ′1 ℓℓ′2 ℓℓ′3 
ℓ ℓ ℓ
 + 6(−1)m′2 Cℓ′ Cℓ′ δℓ′ ℓ′ δm′ −m′ Cℓ (−1)m dn1 ℓ2
ℓ ℓ ℓ3
o
2 3 1 2 1 2 ℓ′
1 2 3
m′m′m′
1 2 3
m′1 m′2 m′3 ℓm
3
m −m m′
ℓ

1 2 3 3
′ X ℓ ℓ ℓ3 ′ X ℓ ℓ ℓ3
+ 6(−1)m3 Cℓ′3 Cℓ′1 δℓ′2 ℓ′3 δm′2 −m′3 Cℓ (−1)m dn1 ℓ2 ℓ ℓ′
o + 6(−1)m1 Cℓ′ Cℓ′ δℓ′ ℓ′ δm′ −m′
1 2 3 1 3 1
Cℓ (−1)m dn1 ℓ2 ℓ′ℓ
o = 0. (53)
1 2
ℓm m −m m′ ℓm m −m m′
1 2

This formula, together with Eq. (45), form a set of equations which completely determines the best unbiased estimator.
We see the complicated structure of these equations. The last three terms come from the 9dd term in the variance and
are not present in the case of the angular spectrum.
 13

From this last equation and using the constraint Eq. (45) we can get the general expression for the Lagrange multipliers.
Thus, we multiply Eq. (53) by the appropriate 3j-symbol and we sum over the three indices m′i . The first term is exactly
ℓi ℓ′j
the constraint and produces a δS . Using the fact that a triple sum over the mi ’s of the squared of a 3j-symbol gives
unity, the second term yields the Lagrange multipliers themselves. Finally, the last three terms vanish: indeed, after
straightforward manipulations one generates a term like
 
ℓ1 ℓ2 ℓ3
X
m′1 ℓ′1 ℓ′1 ℓ′3
dn ℓ ℓ′
o (−1)   (54)
m −m
ℓ
3
0 m′1 m′1 −m′1 0

for the third term of Eq. (53) and analogously for the last two ones. As we mentioned previously, the coefficient d in Eq.
(54) vanishes unless ℓ′3 is even. In this case, recalling the identity (Mollerach et al. 1995)
 
′ ′ ′
X ′ ℓ1 ℓ1 ℓ3 ′ p
(−1)m1   = (−1)ℓ1 2ℓ′1 + 1 δℓ′ 0 , ℓ′3 = even,
3
(55)
m′ m′1 −m′1 0
1

we see that there will only be a non-vanishing term if ℓ′3 = 0. But, the corresponding d is zero because ℓ′3 < 2 and therefore
terms of this kind do not contribute to the Lagrange multipliers. Then, these are given by
ℓi ℓ′j
λℓℓ1′ ℓℓ2′ ℓℓ3′ = −12Cℓ′1 Cℓ′2 Cℓ′3 δS . (56)
1 2 3

Plugging this into Eq. (53), one has

 
ℓ′1 ℓ′2 ℓ′3
 
ℓ1 ℓ2 ℓ3 ℓi ℓ′j ′ X ℓ ℓ ℓ
12Cℓ1 Cℓ2 Cℓ3 d ℓ′ ℓ′ ℓ′ − δS
′ ′ ′ n o   + 6(−1)m2 Cℓ′2 Cℓ′3 δℓ′1 ℓ′2 δm′1 −m′2 Cℓ (−1)m dn1 ℓ2 3ℓ ℓ′ o
′ ′ ′
1 2 3
m′m′m′
m1 m2 m3 ℓm m −m m′
3
1 2 3 3
X X
m′3 ′ m ℓn 1 ℓ2 ℓ3 m′1 ′ m ℓn1 ℓ2 ℓ3
+ 6(−1) Cℓ3 Cℓ1 δℓ2 ℓ3 δm2 −m3
′ ′ ′ ′ ′ Cℓ (−1) d ℓ ℓ ℓ′ + 6(−1) Cℓ1 Cℓ2 δℓ3 ℓ1 δm3 −m1
o ′ ′ ′ ′ ′ Cℓ (−1) d ℓ ℓ ℓ′
o = 0. (57)
1 2
ℓm m −m m′ ℓm m −m m′
1 2

This is the final equation to be solved in order to determine the best unbiased estimator. A solution is
 
ℓ′1 ℓ′2 ℓ′3 ′
ℓ1 ℓ2 ℓ3
dn ℓ′ ℓ′ ℓ′ o =   δ ℓi ℓj . (58)
S
1 2
m′ m′ m′
3 m′1 m′2 m′3
1 2 3

This leads to  
ℓ1 ℓ2 ℓ3 ′ ′ ′
 am1 am2 am3 .
X
EBest (Cℓ1 ℓ2 ℓ3 ) = 
ℓ1 ℓ2 ℓ3 (59)
m′1 m′2 m′3 m′1 m′2 m′3
This is the main result of this subsection.
Given that we now know the best unbiased estimator for Cℓ1 ℓ2 ℓ3 , one can compute its variance, the smallest one amongst
all possible estimator variances. In (Gangui & Martin 2000) we have already calculated it and reads

σE2Best (Cℓ ) = Cℓ1 Cℓ2 Cℓ3 (1 + δℓ1 ℓ2 + δℓ2 ℓ3 + δℓ3 ℓ1 + 2 δℓ1 ℓ2 δℓ2 ℓ3 ). (60)
1 ℓ2 ℓ3

In the same reference a plot of this variance for low order multipoles can also be found. This is what one could dub (the
square of) the ‘bispectrum cosmic variance’ in perfect analogy with σE2Best (Cℓ ) = 2Cℓ2 /(2ℓ + 1), which is (the square of) the
variance of the best unbiased estimator for the angular spectrum, commonly known as the ‘cosmic variance’.
Let us conclude this subsection by comparing our results with those recently appeared in the literature. An estimator
restricted to the diagonal case ℓ1 = ℓ2 = ℓ3 has been proposed in (Ferreira et al. 1998, see also Magueijo 1999 for an
extension of their analysis) for Bℓ ≡ Cℓℓℓ and reads
 −3/2  
1  ℓ ℓ ℓ  X ℓ ℓ ℓ
 am1 am2 am3 .
E(Bℓ ) = 
ℓ ℓ ℓ (61)
2ℓ + 1 0 0 0 m1 m2 m3 m1 m2 m3
 14

In that work, the aim of the authors was not to seek the best estimator, but to use Eq. (61) to analyse the non-Gaussian
features of the 4-yr COBE-DMR data. It is easy to see that their estimator does not satisfy the constraint (45), i.e. the
estimator is biased. This is due to the presence of the overall prefactor in front of the triple sum in Eq. (61). However,
as we have proven above, getting rid of it produces the best unbiased estimator EBest (Cℓ1 ℓ2 ℓ3 ), Eq. (59).

4.2. Best estimator for the third moment B α

We now seek an estimator for B α ≡ am

1 m2 m3 
ℓ 1 aℓ 2 aℓ 3 where, as in the last section, it is convenient to define a collective
ℓ1 ℓ2 ℓ3
index α ≡ { m1 m2 m3
}. This question has already been addressed in (Heavens 1998). As we did with the second
moment, our starting expression for an unbiased cubic estimator E(B α ) of B α will be in the form
Z Z Z
E(B α ) = dΩ1 dΩ2 dΩ3 E α (e1 , e2 , e3 )∆(e1 )∆(e2 )∆(e3 ) (62)

The goal is to find the weight function E α (e1 , e2 , e3 ) that minimizes the variance of the estimator. As above, the quantity
B α is unchanged if we permute arbitrary columns of indices in α: B α = B ᾱ , where for instance ᾱ ≡ { ℓ2
m2
ℓ1
m1
ℓ3
m3
}. E(B α )
has the same properties as B α and then it follows that E(B α ) = E(B ᾱ ) for any column-permutated ᾱ. This implies that
the E α (e1 , e2 , e3 ) satisfies E α (e1 , e2 , e3 ) = E ᾱ (e1 , e2 , e3 ).
Now, is E α (e1 , e2 , e3 ) also symmetric under permutations in the directions e1 , e2 , e3 ? From its definition we cannot
know, for these directions are integrated over in the above defining equation. Unlike the case for the second moment
discussed before, here E α cannot be decomposed into a symmetric and antisymmetric parts. However, we can always
write E α = ESα + something, and show that this last contribution to Eq. (62) vanishes. Therefore, there is no loss of
generality in working with ESα (e1 , e2 , e3 ) ≡ 16 [E α (e1 , e2 , e3 ) + 5 terms ] which is symmetric under arbitrary permutations
of directions ei .
Demanding the estimator E(B α ) to be unbiased, E(B α ) = B α , yields the first constraint equation that the weight



function ESα (e1 , e2 , e3 ) must satisfy

Z Z Z
′ ′
dΩ1 dΩ2 dΩ3 ESα (e1 , e2 , e3 )RR
α
(e1 , e2 , e3 ) = δSαα , (63)

′ m′ m′ m′ ′
α
where Rα (e1 , e2 , e3 ) ≡ Yℓ′ 1 (e1 )Yℓ′ 2 (e2 )Yℓ′ 3 (e3 ) and RR (e1 , e2 , e3 ) is its real part. The form of Rα comes from the
1 2 3

expression of ξ3 , viz. ∆(e1 )∆(e2 )∆(e3 ) = α B α Rα (e1 , e2 , e3 ). The symmetrized Krönecker symbol can be written as

 P
′
1
δSαα ≡ 6 (δℓ1 ℓ1 δm1 m1 δℓ2 ℓ2 δm2 m2 δℓ3 ℓ3 δm3 m3 + 5 terms) as required to comply with the symmetry under permutations in
′ ′ ′ ′ ′ ′

α α′
the columns of α in ES (e1 , e2 , e3 ). In the above equation, RR (e1 , e2 , e3 ) is clearly non-symmetric under a permutation
of directions e1 , e2 , e3 . However, as above, we can define a symmetrized combination RSα (e1 , e2 , e3 ) ≡ 16 [RR α
(e1 , e2 , e3 ) +
α
5 terms] (12 terms). Symmetrizing either in directions or in the columns of α in RR (e1 , e2 , e3 ) yields exactly the same
RSα (e1 , e2 , e3 ). In the last equation and in what follows the weight function ESα is real for reasons similar to the ones
exposed around Eq. (28) in the last section.
Proceeding as in §3 and using Eq. (9), one gets
Z Z h i
2
σE(B α) = dΩ1 . . . dΩ6 ESα (e1 , e2 , e3 )ESα (e4 , e5 , e6 )ξ(e1 ·e4 ) 6ξ(e2 ·e5 )ξ(e3 ·e6 ) + 9ξ(e2 ·e3 )ξ(e5 ·e6 ) , (64)

where, utilizing the symmetry of the coefficients ESα (ei , ej , ek ) under e-direction permutations, only two types of ξ products
remain: first type, six terms where all the three ξ’s mix directions of the first and second ESα ’s and, second type, nine
terms where only one ξ [in the above equation, ξ(e1 ·e4 )] does it.
To minimize the remaining variance under the constraint (63) we introduce a set of Lagrange multipliers λα and write
h X Z Z Z i
2 α α′ ′
δ σE(Bα ) − λα′ dΩ1 dΩ2 dΩ3 ESα (e1 , e2 , e3 )RR (e1 , e2 , e3 ) − δSαα = 0. (65)
α′
 15

As already noted for the second moment, the symmetries of the weight function must be respected in the variation;
hence we have
δESα (e1 , e2 , e3 ) 1 h δ(e1 ·ei − 1) δ(e2 ·ej − 1) δ(e3 ·ek − 1) i
= + 5 terms δSαβ . (66)
δESβ (ei , ej , ek ) 6 2π 2π 2π
Now, we vary Eq. (65) and, after some algebra, we come up with
1n
Z Z Z h i o X
α′
2 dΩ4 dΩ5 dΩ6 ξ(ei ·e4 ) 6ξ(ej ·e5 )ξ(ek ·e6 ) + 9ξ(ej ·ek )ξ(e5 ·e6 ) + 5 terms ESα (e4 , e5 , e6 ) = λα
α′ RS (ei , ej , ek ),
6
α′
(67)
′
expression which is symmetric in the arbitrary directions ei , ej , ek as it should (note the presence of RSα ).
We aim at getting an explicit expression for ESα (e4 , e5 , e6 ). A glance at the previous equation shows that we need to
multiply both sides of it by the inverse of the correlation function ξ −1 defined in Appendix. Concretely, we multiply Eq.
(67) by ξ −1 (ei ·ei′ )ξ −1 (ej ·ej ′ )ξ −1 (ek ·ek′ ) and integrate over directions ei , ej , ek ; we get

6 ESα (ei′ , ej ′ , ek′ )

(2π)2
ZZ h i
+9 dΩ5 dΩ6 ξ(e5 ·e6 ) ξ −1 (ei′ ·ej ′ )ESα (ek′ , e5 , e6 ) + ξ −1 (ej ′ ·ek′ )ESα (ei′ , e5 , e6 ) + ξ −1 (ek′ ·ei′ )ESα (ej ′ , e5 , e6 )
3
1
ZZZ X
α′
= dΩi dΩj dΩk ξ −1 (ei ·ei′ )ξ −1 (ej ·ej ′ )ξ −1 (ek ·ek′ ) λαα′ RS (ei , ej , ek ). (68)
2 ′α
R
As before, the 2π factors in the left hand side come from operations like 2πf (ei ) = dΩk δ(ei · ek − 1)f (ek ), for an
arbitrary function f (ek ). However, we don’t have an explicit expression for ESα yet. We see that expressions of the type
ξ(e5 ·e6 )ξ −1
(− · − )ESα (− , e5 , e6 )
are the ones that prevent us from isolating the weight function. To deal with this, it is
dΩ5 dΩ6 ESα (− , e5 , e6 )ξ(e5 ·e6 ). To reach this goal, we multiply both sides of
RR
convenient to construct the combination
Eq. (68) by ξ(ej ′ ·ek′ ) and integrate over directions ej ′ and ek′ . This operation produces a divergence in the left hand
side of this equation in a form of a Dirac function ‘δ(0)’. In the continuous case, all methods lead to this unavoidable
problem and although it has already appeared in the literature (Heavens 1998), it has never been treated so far. That this
divergence is a mathematical artifact we can see from the fact that, in practice, we never deal with an ideal experiment:
the problem is solved when we take into account the fact that each different experimental setting is limited by a finite
angular resolution. This is usually quantified in terms of an ℓ-dependent window function Wℓ (the circularly symmetric
pattern of the observation beam in ℓ-space), although more involved scanning techniques are also employed (White &
Srednicki 1995, Knox 1999). Then, only a finite number of multipoles will effectively contribute to the correlation function
and, as a result, the above mentioned divergence is regularized; indeed, we have
X 2ℓ + 1
ξ(ei ·ej ) = Cℓ Wℓ2 Pℓ (ei ·ej ), (69)
4π
ℓ

which, upon using the expression for ξ −1 given in the Appendix, leads to the quantity
Z X
Aik ≡ dΩj ξ(ei ·ej )ξ −1 (ej ·ek ) = 2π Wℓ2 Yℓm (ei )Yℓm ∗ (ek ). (70)
ℓm

In particular, the previous divergence ‘δ(0)’ now becomes

X 2ℓ + 1
A ≡ Aii = Wℓ2 . (71)
2
ℓ

This is a more realistic and finite object to work with in the case where two directions on the microwave sky coincide for
a given experience. Notice that for an ideal experimental setting in which the window function Wℓ → 1, or equivalently
the beam σ → 0 in the case of a Gaussian profile, A blows up. Since the terms of the type 9ξξ in Eq. (64) are not
16

100
75
50
60
25
0
40
20
40 20
60

Fig. 2.— Values of Aik (as defined in the text) with fixed direction ei at the center of the plot and direction ek scanning a square of
side 90◦ . We show the result for the COBE-DMR window function. The full width at half maximum of the plot is roughly 10◦ , of the same
order as the resolution of the COBE-DMR detector. We can then expect the plotted ‘spike’ to select just one experimental pixel on the sky
map, making the relation Aik ≈ δ(ei ·ek − 1) a good approximation. Note that higher resolution experiments will yield a more peaked curve,
although one would expect the goodness of the approximation to remain roughly the same.

present in the case of the two-point correlators, this problem did not appear there. From now on, strictly speaking, all
expressions should incorporate the window function. However, in what follows and for computational convenience, we
will keep Aik ≈ δ(ei ·ek − 1) for ei 6= ek , a good approximation as we can see from Fig. 2.
dΩ5 dΩ6 ESα (− , e5 , e6 )ξ(e5 · e6 ) in the
RR
Endowed now with the above regularization method, we present the term
following form
D X α
ZZ ZZZ
′
dΩ5 dΩ6 ESα (e4 , e5 , e6 )ξ(e5 ·e6 ) = 2
λα′ dΩi′ dΩj ′ dΩk′ ξ −1 (e4 ·ei′ )ξ −1 (ej ′ ·ek′ )RSα (ei′ , ej ′ , ek′ ) (72)
24π ′α

2
with D ≡ 24π 3 [6 + 9 (2π)
3 (2 × 2π + 4πA)]
−1
where, as expected, the factor A appears explicitly. Now, we just replace the
six terms with prefactor 9ξξ in the left hand side of Eq. (67) with Eq. (72) resulting in
ZZZ
dΩ4 dΩ5 dΩ6 ξ(ei ·e4 )ξ(ej ·e5 )ξ(ek ·e6 )ESα (e4 , e5 , e6 ) (73)
1 X α n α′
= λ ′ RS (ei , ej , ek )
12 ′ α
α
D
ZZ h ′ io
′ ′
− dΩj ′ dΩk′ ξ −1 (ej ′ ·ek′ ) RSα (ei , ej ′ , ek′ )ξ(ej ·ek ) + RSα (ej , ej ′ , ek′ )ξ(ek ·ei ) + RSα (ek , ej ′ , ek′ )ξ(ei ·ej ) .
2π
This contains just one appearance of E α ; to get the weight function explicitly, we only need to multiply the equation by
ξ −1 (ei ·e1 )ξ −1 (ej ·e2 )ξ −1 (ek ·e3 ) and integrate over directions ei , ej , ek . Finally, we get the expression for ESα in terms of
the Lagrange multipliers λα
1 X nZ Z Z ′
ESα (e1 , e2 , e3 ) = λα
dΩi dΩj dΩk ξ −1 (ei ·e1 )ξ −1 (ej ·e2 )ξ −1 (ek ·e3 )RSα (ei , ej , ek ) (74)
12 × (2π)3 ′ α
′

α
ZZZ h io
′
−D dΩi dΩj dΩk RSα (ei , ej , ek )ξ −1 (ej ·ek ) ξ −1 (ei ·e1 )ξ −1 (e2 ·e3 ) + ξ −1 (ei ·e2 )ξ −1 (e3 ·e1 ) + ξ −1 (ei ·e3 )ξ −1 (e1 ·e2 )

Reached this point, we have an explicit expression for ESα , but still dependent on the Lagrange multipliers. This equation is
well defined as it contains the renormalized quantity A. Within a particular experiment with a given resolution, the value
A takes depends on what one means by two coincident directions. For example, for the COBE-DMR window-function
 17

specification (a Gaussian beam with dispersion σ = 3◦.2) this yields roughly A ≈ 158.5, including the quadrupole. It is not
difficult to extend this to other scanning techniques. The previous equation is the analogue of our Eq. (34) corresponding
to the second moment and also to Eq. (21) of (Heavens 1998). In that article, a similar analysis is done but for the
discretized CMB sky. Remark that no divergence appears in his case. Indeed, all relevant quantities are finite when
evaluated for two directions pointing towards the same pixel. Our prefactor D corresponds to 3/(2 + 3N ) in that paper,
where N represents the number of pixels in the map. Clearly, N → ∞ when the pixel size goes to zero, as well as
A → ∞ when the window function Wℓ → 1. At this intermediate step our corresponding expressions need not coincide
because both depend on the particular regularization scheme used (be it discretization or usage of a window function).
Despite appearances, we will show below that the final expression for the best unbiased estimator does not depend on
these schemes. This cannot be inferred from Eq. (74) because we still need to remove the Lagrange multiplier. Unlike
what was done in (Heavens 1998), we now proceed further and express the weight function explicitly. Hence, we multiply
α ′′ ℓ′′ ℓ′′ ℓ′′
both sides of Eq. (74) by RR (e1 , e2 , e3 ), where α′′ ≡ { 1
m′′
2
m′′
3
m′′
}, then integrate over directions e1 , e2 and e3 and,
1 2 3

upon using the constraint equation (63), we get

′′ 1 1 nh ′′ ′′ ′′
i o
12δSαα = (−1)m1 +m2 +m3 λα n ′′ ′′ ′′ o + λα n ′′ ′′ ′′ o + 5 additional permutations in index α′′
6 Cℓ′′1 Cℓ′′2 Cℓ′′3 ℓ
1
ℓ
2
ℓ
3
ℓ
1
ℓ
2
ℓ
3
−m′′ −m′′ −m′′ m′′ m′′ m′′
1 2 3 1 2 3
("
m′′ m
D δℓ1 ℓ2 δm1 −m2 (−1)
′′ ′′ ′′ ′′ 2 X (−1)  ′′ ′′
− (−1)m3 λα n
ℓ′′
o + λα
n
ℓ ℓ ℓ′′
o + (−1)m3 λα
n ′′ o + λα
n ′′ o
3 Cℓ′′2 Cℓ′′3 Cℓ ℓ ℓ
3 3
ℓ
3
ℓ ℓ ℓ
3
ℓ ℓ
ℓm m−m−m ′′
3
−mmm ′′
3
′′
−m m−m
3
′′
m −mm
3
#    1 → 3 )
′′
 1 → 2
+ (−1)m3 λα n ′′
ℓ ℓ ℓ
o + λα n
ℓ ℓ ′′
ℓ
o + 2→3 + 2→1 . (75)
3 3 3 → 1 3 →2
m−m′′ −m −mm′′ m
3 3

Eq. (75) is the final algebraic equation that the multipliers must satisfy. For fixed α, a natural way to proceed would
be to get an explicit expression for λα
α′ . Another way to solve the problem goes on a line analogous to the case of the

two-point correlators: we just need to identify the complicated combination of Lagrange multipliers in the right hand side
of Eq. (75) with the one in the right hand side of Eq. (73) [or, equivalently, Eq. (74)]. In order to do that, we now
′′
α
multiply both sides of Eq. (75) by RR (ei , ej , ek ), perform the six sums over the indices in α′′ and we end up with
X ′′ ′′
RSα (ei , ej , ek )Cℓ1 Cℓ2 Cℓ3 = RRα
(ei , ej , ek )Cℓ′′1 Cℓ′′2 Cℓ′′3 δSαα
α′′
1 X α α′ D X X (−1)M h m i
= λα′ RS (ei , ej , ek ) − Yℓ (ei )ξ(ej ·ek ) + Yℓm (ej )ξ(ek ·ei ) + Yℓm (ek )ξ(ei ·ej )
12 ′ 36 CL
α LM ℓm
h i
× λαn
ℓ L L
o + (−1)m λα n
ℓ L L
o + λα
n
L ℓ L
o + (−1)m λα n
L ℓ L
o + λαn
L L ℓ
o + (−1)m λα n
L L ℓ
o . (76)
mM −M −m−M M −M mM M −m−M M −M m −M M −m

This is the equivalent of Eq. (35). The aim now is to show that the combination of Lagrange multipliers in the right hand
side of the previous equation is precisely the one which appears in the right hand side of Eq. (73). In the latter, let us
′
express both RSα and ξ −1 in the second term in the right hand side in terms of spherical harmonics. After some algebra
we get
ZZZ
dΩ4 dΩ5 dΩ6 ξ(ei ·e4 )ξ(ej ·e5 )ξ(ek ·e6 )ESα (e4 , e5 , e6 )

1 X α α′ D X X (−1)M h m i
= λα′ RS (ei , ej , ek ) − Yℓ (ei )ξ(ej ·ek ) + Yℓm (ej )ξ(ek ·ei ) + Yℓm (ek )ξ(ei ·ej )
12 ′ 36 CL
α LM ℓm
h i
× λαn
ℓ L L
o + (−1)m λα n
ℓ L L
o + λα
n
L ℓ L
o + (−1)m λα n
L ℓ L
o + λαn
L L ℓ
o + (−1)m λα n
L L ℓ
o , (77)
mM −M −m−M M −M mM M −m−M M −M m −M M −m

which, as advertised, yields the same combination of Lagrange multipliers of Eq. (76). Then, putting the last two
equations together, multiplying by ξ −1 and integrating three times, we finally get the weight function associated to the
18

best unbiased estimator

α
ES,Best (ei , ej , ek ) = RSα (ei , ej , ek ), (78)

which implies that the best unbiased estimator itself is given by

1  m1 m2 m3 
EBest (B α ) = aℓ 1 aℓ 2 aℓ 3 + am
ℓ1
1 ∗ m2 ∗ m3 ∗
a ℓ2 a ℓ3 . (79)
2
This is the final answer and it is a new result. Let us make a few remarks. Firstly, this does not depend on A which
shows that Eq. (79) is independent of the regularization scheme used. Secondly, Luo (1994) used the following complex
unbiased estimator: E(B α ) = am 1 m2 m3
ℓ1 aℓ2 aℓ3 , although he did not claim it to be the best one. Thirdly, as for the two-point
correlators, one cannot use Eq. (79) in order to infer EBest (Cℓ1 ℓ2 ℓ2 ). Indeed, promoting the equation
 

m1 m2 m3  ℓ1 ℓ2 ℓ3
aℓ 1 aℓ 2 aℓ 3 =   C ℓ1 ℓ2 ℓ3 , (80)
m1 m2 m3

valid for the mean values of the estimators, to the following equation
 
1   ℓ1 ℓ2 ℓ3
EBest (B α ) = am 1 m2 m3 m1 ∗ m2 ∗ m3 ∗
ℓ 1 aℓ 2 aℓ 3 + aℓ 1 aℓ 2 aℓ 3 = EBest (Cℓ1 ℓ2 ℓ3 )   (false), (81)
2 m1 m2 m3

valid for the estimators themselves, is an unjustified step. If, nevertheless, we used this false relation we would get
   
ℓ 1 ℓ 2 ℓ 3 ℓ ℓ ℓ
 1 2 3 ′ ′ ′
 am1 am2 am3 (false),
X
EBest (B α ) =  
ℓ 1 ℓ 2 ℓ 3
(82)
m1 m2 m3 m′1 m′2 m′3 m′1 m′2 m′3
 
which cannot be cast into 21 am 1 m2 m3
a a
ℓ1 ℓ2 ℓ3 + a m1 ∗ m2 ∗ m3 ∗
ℓ1 a ℓ2 a ℓ3 . So, like for the two-point correlators, we see that one cannot
infer the EBest (B α ) from EBest (Cℓ1 ℓ2 ℓ3 ) and vice versa.
Endowed now with the best unbiased estimator, one can compute its variance squared, the ‘third-moment cosmic
variance’, which reads
1n h i h i
σE2Best (Bα ) = Cℓ1 Cℓ2 Cℓ3 1 + δm1 0 δm2 0 δm3 0 + Cℓ31 δℓ1 ℓ2 ℓ3 8δm1 0 δm2 0 δm3 0 + 2(δm1 m3 + δm1 −m3 )(δm1 m2 + δm1 −m2 )
2 h i h i
+ Cℓ1 Cℓ22 δℓ2 ℓ3 2δm1 0 δm2 −m3 + δm2 m3 + δm2 −m3 + Cℓ2 Cℓ23 δℓ3 ℓ1 2δm2 0 δm3 −m1 + δm3 m1 + δm3 −m1
h io
+ Cℓ3 Cℓ21 δℓ1 ℓ2 2δm3 0 δm1 −m2 + δm1 m2 + δm1 −m2 . (83)

For example, from this we can now compute the cosmic variance for the third-moment estimator EBest (B β ) with β ≡
ℓ ℓ ℓ
{ m m −2m
} where ℓ = even and m 6= 0. This particular case is often treated in the literature, see e.g. (Luo 1994, Heavens
1998). We find σE2Best (Bβ ) = Cℓ3 whereas the variance of the estimator used in (Luo 1994) yields σE(B
2 3
β ) = 2 Cℓ . Another

} where |mi | 6= |mj | for any i, j. With this choice we get σE2Best (Bγ ) = Cℓ3 /2
ℓ ℓ ℓ
example comes from taking γ ≡ { m1 m2 m3
2 3
whereas Luo (1994) obtains σE(B γ ) = Cℓ . Note that in both examples the results differ by a 1/2 factor. This can be

traced back to the form of the best estimator in Eq. (79). The variance computed in (Luo 1994) is consistent with
his choice of the estimator. However, unlike what is stated in that paper, this variance does not deserve the name
‘cosmic’ because, as we saw above, this estimator is not the best one. As expected, the variance of the best estimator
is smaller than the variance computed in his article. Since we have now the correct expression for the cosmic variance,
2 2 2
its numerical value should be re-estimated. To be specific, let us take β ≡ { 1 1 −2
}. The cosmic variance is then
3/2 3/2
σEBest (Bβ ) = C2 = (4π/5) Q3rms−PS /T03 ≈ 1.3 × 10 −15
where we used T0 = 2.7 K and Qrms−PS = 18.7µK (Bunn &
White 1997). Although this figure is close to Luo’s result (1.4 × 10−15 ) cited after equation (32) in (Heavens 1998), this
does not imply that the two variances are not different by a factor 1/2, as we have just seen. This might probably be due
to a difference in the quadrupole normalizations.
 19

5. CONCLUSIONS

Optimized analyses of CMB datasets involve the use of appropriate methods in order to reduce the various uncertainties.
In particular, the theoretical error bars due to the cosmic variance can be minimized by working with the method of the
best unbiased estimators. In this article, we have applied this technique for the study of CMB non-Gaussian features.
These are often characterized by means of the third moment for the am
ℓ ’s or by the angular bispectrum Cℓ1 ℓ2 ℓ3 . We have
found the best unbiased estimators in both cases. These are the quantities that should be used in future data analyses and
would be important for upcoming megapixel experiments (Tegmark 1997, Borrill 1999) like MAP2 and Planck Surveyor3.
In addition to this, we have displayed both the angular bispectrum and the third moment cosmic variances, the smallest
possible uncertainties attached to the bispectrum and the third moment, which would be present in any ideal experiment
when all other sources of noise have been removed.

ACKNOWLEDGMENTS

A.G is member of CONICET, Argentina.

6. APPENDIX: THE INVERSE TWO-POINT CORRELATION FUNCTION

In this Appendix, we derive the exact expression of the inverse of the two-point correlation function ξ −1 , defined
according to Z
dΩj ξ(ei ·ej )ξ −1 (ej ·ek ) ≡ δ(ei ·ek − 1). (84)

As usual, one expands the two-point correlation function on the basis of the Legendre polynomials
X 2ℓ + 1
ξ(ei ·ej ) = Cℓ Pℓ (ei ·ej ). (85)
4π
ℓ

In general, ξ −1 can be expanded on the spherical harmonics basis as follows

XX ′
ξ −1 (ei ·ej ) = bℓℓ′ mm′ Yℓm (ei )Yℓm
′
∗
(ej ). (86)
ℓm ℓ′ m′

Our aim now is to determine the coefficients bℓℓ′ mm′ . Using Eqns. (85) and (86), the completeness relation for the
Legendre polynomials together with the addition theorem of spherical harmonics in Eq. (84), one gets
XX ′ X 1
Cℓ bℓℓ′ mm′ Yℓm (ei )Yℓm
′
∗
(ek ) = (ℓ′′ + )Pℓ′′ (ei · ek ). (87)
2
ℓm ℓ′ m′ ℓ′′

The coefficients bℓℓ′ mm′ are easy to read off and one obtains

2π
bℓℓ′ mm′ = δℓℓ′ δmm′ , (88)
Cℓ
from which one deduces
X ℓ + 1/2
ξ −1 (ei ·ej ) = Pℓ (ei · ej ). (89)
Cℓ
ℓ

This is the expression used in the main text.

2 http://map.gsfc.nasa.gov/

3 http://astro.estec.esa.nl/Planck/
20

REFERENCES
Albrecht, A. & Steinhardt, P. J. 1982, Phys. Rev. Lett. 48, 1220.
Avelino, P .P. et al. 1998, Astrophys. J., 507, L101.
Banday, A. J., Zaroubi, S. & Górski, K. M, astro-ph/9908070.
Bardeen, J. M., Steinhardt, P. J. & Turner, M. S. 1983, Phys. Rev. D 28, 679.
Bernardeau, F. 1997, Astronomy and Astrophysics, 324, 15.
Bond, J. R. & Efstathiou, G. 1987, MNRAS 226, 655.
Borrill, J. 1999, astro-ph/9911389.
Bouchet, F. & Gispert, R. 1999, astro-ph/9903176.
Bouchet, F. R., Bennett, D. P. & Stebbins, A. 1988, Nature 335, 410.
Brandenberger, R. H. & Hill, C. T. 1986, Phys. Lett. B 179, 30.
Bromley, B. & Tegmark, M. Astrophys. J. Lett. in press (astro-ph/9904254).
Bunn, E. & Scott, D. 1999, astro-ph/9906044.
Bunn, E. F. & White, M. 1997, Astrophys. J. 480, 6.
Coles, P. 1989, MNRAS, 234, 509.
Contaldi, C. R., Bean, R. & Magueijo, J. 1999, Phys. Lett. B 468, 189.
Cooray, A. & Hu, W. 1999, astro-ph/9910397.
Falk, T., Rangarajan, R. & Srednicki, M. 1993, Astrophys. J. 403, L1.
Ferreira, P. & Magueijo, J. 1997, Phys. Rev. D 55, 3358.
Ferreira, P. G., Magueijo, J. & Górksi, K. M. 1998, Astrophys. J. 503, L1.
Gangui, A. 1994, Phys. Rev. D. 50, 3684.
Gangui, A., Lucchin, F., Matarrese, S. & Mollerach, S. 1994, Astrophys. J. 430, 447.
Gangui, A. & Martin, J. 2000, Mon. Not. R. Astron. Soc. 313, 323. astro-ph/9908009.
Gangui, A. & Mollerach, S. 1996, Phys. Rev. D 54, 4750.
Gangui, A. & Perivolaropoulos, L. 1995, Astrophys. J. 447, 1.
Goldberg, D. M. & Spergel, D. N. 1999, Phys. Rev D 59, 103002.
Górksi, K. M. 1994, Astrophys. J. 430, L85.
Gott, J. R. et al. 1990, Astrophys. J. 352, 1.
Grishchuk, L. P. & Sidorov, Y. 1990, Phys. Rev. D 42, 3413.
Grishchuk, L. P. 1993, Phys. Rev. D 48, 3513.
Grishchuk, L. P. & Martin, J. 1997, Phys. Rev. D 56, 1924.
Guth, A. 1981, Phys. Rev. D 23, 347.
Halliwell, J. J. 1989, in Quantum Cosmology and Baby Universes, Proceedings of the 1989 Jerusalem Winter School for Theoretical Physics,
Eds. Coleman S. et al., World Scientific, Singapore.
Hawking, S. W. 1982, Phys. Lett. B 115, 295.
Heavens, A. 1998, MNRAS 299, 805.
Heavens, A. & Sheth, R. 1999, astro-ph/9904307.
Hinshaw, G. et al. 1994, Astrophys. J. 431, 431.
Hinshaw, G. et al. 1995, Astrophys. J. 446, L67.
Hobson, M. et al. 1998, astro-ph/9810200.
Knox, L. 1999, astro-ph/9902046.
Kogut, A. et al. 1995, Astrophys. J. 439, L29.
Kogut, A. et al. 1996, Astrophys. J. 464, L29.
Linde, A. 1983, Phys. Lett. B 108, 389.
Linde, A. 1983, Phys. Lett. B 129, 177.
Linde, A. & Mukhanov, V. 1997, Phys. Rev. D 56, 535.
Luo, X. 1994, Astrophys. J. 427, L71.
Luo, X. & Schramm, D. N. 1993, Phys. Rev. Lett. 71, 1124.
Magueijo, J. 1999, astro-ph/9911334.
Martin, J., Riazuelo, A. & Sakellariadou, M. 1999, astro-ph/9904167, to appear in Phys. Rev. D.
Martin, J. & Schwarz, D. J. 1998, Phys. Rev. D 57, 3302.
Martin, J. & Schwarz, D. J. 1999, astro-ph/9911225.
Mollerach, S., Gangui, A., Lucchin, F. & Matarrese, S. 1995, Astrophys. J. 453, 1.
Mukhanov, V. F. & Chibisov, G. 1981, JETP lett. 33, 532.
Mukhanov, V. F., Feldman, H. A. & Brandenberger, R. H. 1992, Phys. Rep. 215, 203.
Munshi, D., Souradeep, T. & Starobinsky, A. A. 1995, Astrophys. J. 454, 552.
Pando, J., Vallas-Gabaud, D. & Fang, L. 1998, Phys. Rev. Lett. 79, 1611.
Popa, L. 1998, astro-ph/9806086.
Scaramella, R. & Vittorio, N. 1991, Astrophys. J. 375, 439.
Scott, D. 1999, astro-ph/9912038.
Smoot, G. et al. 1994, Astrophys. J. 437, 1.
Spergel, D. N & Goldberg, D. M. 1999, Phys. Rev. D 59, 103001.
Srednicki, M. 1993, Astrophys. J. 416, L1.
Starobinsky, A. A. 1982, Phys. Lett. B 117, 175.
Tegmark, M. 1997, Astrophys. J. 480, L87.
Tegmark, M. et al. 1999, astro-ph/9905257.
Wang, L. & Kamionkowski, M. 2000, PRD, astro-ph/9907431.
White, M. & Srednicki, M. 1995, Astrophys. J. 443, 6.
Winitzki, S. 1998, astro-ph/9806105.
Winitzki, S. & Kosowsky, A. 1998, New Astronomy 3, 75.
Zaldarriaga, M. 1999, astro-ph/9910498.